Meridian Line's charter-desk assistant is supposed to answer only from retrieved charterparty clauses and cite the clause it used. During evaluation, several responses give an answer that sounds correct and matches what the reference answer says, but the response includes no clause citation at all and the retrieved context that turn contained nothing relevant. What does this specific pattern indicate, and what should the evaluation flag?

An answer matching the reference but arriving with no citation and no relevant retrieved context is a strong signal that the model answered from its own general pretraining knowledge rather than from Meridian Line's actual charterparty documents, which is exactly the ungrounded, uncited failure mode this evaluation should catch, because a RAG system that happens to be right by coincidence is not a system that reliably grounds its answers, and the same failure on a different query would produce a confidently wrong, uncited answer with no way to trace where it came from. Treating this as a cost issue misreads the signal entirely, token count has no bearing on whether a response is grounded in retrieved content, an uncited answer can be short or long regardless of its grounding. Dismissing it as a correct answer with only a cosmetic citation gap misses the actual point of requiring citations, the citation is the evidence that the answer came from the governed source material rather than the model's general knowledge, and its absence combined with irrelevant retrieval is a substantive grounding failure, not a formatting nitpick. Framing it as a latency issue has no basis, nothing in the scenario suggests speed pressure caused the model to omit a citation, and citation presence is not a function of response time. Exam framing: matching the reference answer's wording is not sufficient evidence of correctness, checking for grounding and citation against actually retrieved context is a distinct and necessary evaluation signal.